Reproduce “Generalist Graph Anomaly Detection via Prototype-Based Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yiming Xu, Zihan Chen, Zhen Peng, Song Wang, Bin Shi, Bo Dong, Chao Shen
ProMoS is an unsupervised, zero-shot graph anomaly detector that transfers normality knowledge across heterogeneous graphs through prototype-based distillation. Using a frozen self-supervised teacher, shared and personalized student branches, and reliability-weighted prototypes, it achieved the top AUROC on 9 of 11 unseen test graphs and improved average AUROC by 14.12 percentage points over the reported strongest baseline.
Key insights
- Prototype-level soft-label distillation is more transferable across graph domains than direct instance-level embedding matching, because it is less dependent on feature vocabulary, scale, and topology.
- The shared branch captures broadly recurring normality, whereas the personalized branch models local or domain-specific regularities; their combination is supported by variance reduction from ensemble averaging.
- The largest ablation degradation comes from removing prototype-guided distillation, with average AUROC falling from 75.46 to 56.32 across the reported six-dataset study, highlighting the central role of semantic alignment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.